OpenAI's GPT-6.1 Sol matches Astra at one-fifth cost

OpenAI describes GPT-6 Astra as its highest-performing model and positions GPT-6.1 Sol for users who want to handle important work more frequently, or for API users building and running applications at scale. The gap between the two is framed less around raw capability than around cost: on the DeepSWE v1.1 coding and GDP.pdf document-reading tests, OpenAI says GPT-6.1 Sol reaches GPT-6 Astra-equivalent performance at one-fifth the cost per task. On AutomationBench, which measures multi-step workflow execution, OpenAI says the medium-thinking GPT-6.1 Sol beats Anthropic's Claude Opus 5.5 at about one-third the cost, and on OSWorld 2.0, which measures computer-operation ability, it says the score gap with GPT-6 Astra narrows to 2.1 points at about one-seventh the cost per task.

The timing is notable: GPT-6.1 Sol arrives just one week after GPT-6 Sol, and OpenAI reports that in its own alignment evaluations GPT-6.1 Sol improves substantially over GPT-6 Sol and approaches GPT-6 Astra. That framing sits alongside reports from The Wall Street Journal and Reuters that OpenAI withheld the launch of GPT-6.1 Astra, planned for October, over safety concerns.

For developers and businesses, the practical question is whether OpenAI's own benchmark figures translate into day-to-day work, since the comparisons rest on OpenAI's published evaluations. The models reach the same plans and the same API, so the outcome likely hinges on whether teams see the stated cost-per-task advantage in their own workloads rather than on the benchmark scores alone.